So this resolves positively every year this happens? Not just the first? Because i think people are betting as if its just the first

@digory Just the first yeah

updated the description, hopefully is clearer

I imagine at some point providers will just add a standard monitoring output to their hosted models of “someone is asking me to do something sus, flag for review”. Does that qualify as snitching?

@pm if the ai is told in its prompt to flag sus ppl for review, that is snitching. if the ai assistant says nothin but there's an automated monitoring system beyond the ai itself that flags bad words or whatever, outside of the ai assistant's control, that would not be sufficient for a positive resolution. does that answer your question?

@Bayesian almost. I’m imaging this being baked into the model, as another output alongside the next token selector. So not like an API that the language part would call “voluntarily”, just reflexively setting the output from it “knowing” something bad is up.

From the Claude 4 system card:

@Haiku claude can just do things

forgot to mention that in the market description, ty for sharing

@Bayesian I thought this might have been the inspiration; otherwise, that would have been a hilarious coincidence.

like an agent autonomously, deliberately telling authorities about something? or would "someone uses chatgpt to plan crime, openai gets subpoenaed, they get caught using this info'

@Robincvgr I added

As in, an agent autonomously, deliberately telling authorities or others about something the user did or is planning to do, in a way that leads to that user's arrest